Presentation
75. SemanticSRJudge: Region-Aware Super-Resolution Diagnosis via Semantic Drift Heatmaps and Perceptual Evaluation
DescriptionWe propose SemanticSRJudge, a training-free pipeline that routes a vision-language model to the most distorted regions of a super-resolved image (localised via DINOv2) and judges them at native resolution—rather than scoring the full image, where micro-failures are diluted by intact context. Our key novelty is region-conditioned, dimensional judging: diagnosing what failed and where, instead of collapsing quality into a single scalar. Through a large-scale study spanning four SR architectures, four datasets, both 2× and 4× scales, and thousands of judge calls, we establish three findings. (1) Region conditioning deflates overly optimistic full-image scores on texture and upsampling quality, where small failures hide. (2) We uncover a content-dependent prior-selection rule no scalar metric captures: diffusion priors (SUPIR) dominate at 4× on faces, animals, and indoor scenes, yet hallucinate plausible-but-wrong detail on text and patterns, losing to Real-ESRGAN+. (3) Region conditioning significantly improves alignment with human preference.
Event Type
Poster
TimeSunday, 19 July 20269:00am - 5:30pm PDT
LocationWest Hall Lobby
Session TimeSunday, 19 July 20269:00am - 5:30pm PDTMonday, 20 July 20269:00am - 5:30pm PDTTuesday, 21 July 20269:00am - 5:30pm PDTWednesday, 22 July 20269:00am - 5:30pm PDTThursday, 23 July 20269:00am - 5:30pm PDT
LocationWest Hall Lobby
Full Conference Supporter
Full Conference
Experience
